Finds where an origin-centered ray going in direction trg intersects a triangle. Args: tri: 3 X 3 vertex locations. tri[0, :] is 0th

verts: N x 3 vert2kp: K x N For each keypoint, visualize its weights on each vertex. Base color is white, pick a color for each kp.

Input: img_pred, img_gt: B x 3 x H x W mask_pred, mask_gt: B x H x W

Pred and target are Variables. If normalize is on, scales images between [-1, 1] Assumes the inputs are in range [0, 1].

Initialize virtual trackball control. initial : quaternion or rotation matrix

Try import all public attributes from module into global namespace. Existing attributes with name clashes are renamed with prefix. Attributes

Return angle between vectors. If directed is False, the input vectors are interpreted as undirected axes, i.e. the maximum angle is pi/2.

Just L'g = Lg Args: grad_out: B x N x 3 Returns: grad_vertices: B x N x 3

cam_* are B x 7, [sc, tx, ty, quat] Losses are in similar magnitude so one margin is ok.

Return matrix to obtain normalized device coordinates from frustum. The frustum bounds are axis-aligned along x (left, right), y (bottom, top

Return transformation matrix from sequence of transformations. This is the inverse of the decompose_matrix function. Sequence of transformat

Computes distance transform of mask.

Computes barrier distance transform of mask.

Returns a list: [A, B, C, D] the 4 vertices for each edge.

For this mesh, pre-computes the UV coordinates for F x T x T points. Returns F x T x T x 2

Return concatenation of series of transformation matrices. >>> M = numpy.random.rand(16).reshape((4, 4)) - 0.5 >>> numpy.allclose(M, concaten

Return state of constrain to axis mode.

Crops a region from the image corresponding to the bbox. If some regions specified go outside the image boundaries, the pixel values are set

Cross product of va by vb. Args: qa: B X N X 3 vectors qb: B X N X 3 vectors Returns: q_mult: B X N X 3 vectors

Simple 3D encoder with nlayers. Args: nlayers: number of decoder layers nz_shape: number of bottleneck nc_input: num

Return sequence of transformations from transformation matrix. matrix : array_like Non-degenerative homogeneous transformation matrix

l2 norm on V = B x N x 3

Set initial cursor window coordinates and pick constrain-axis.

Update current cursor window coordinates.

Simple 3D encoder with nlayers. Args: nlayers: number of encoder layers use_bn: whether to use batch_norm nc_input:

Input is K x N Each column is a prob of vertices being the one for k-th keypoint. We want this to be sparse = low entropy.

Return Euler angles from quaternion for specified axis sequence. >>> angles = euler_from_quaternion([0.99810947, 0.06146124, 0, 0]) >>> numpy

Return 4x4 identity/unit matrix. >>> I = identity_matrix() >>> numpy.allclose(I, numpy.dot(I, I)) True >>> numpy.sum(I), numpy.trace(

Print methods and doc strings. Takes module, class, list, dictionary, or string.

Return inverse of square transformation matrix. >>> M0 = random_rotation_matrix() >>> M1 = inverse_matrix(M0.T) >>> numpy.allclose(M1, nu

Return True if two quaternions are equal.

Return True if two matrices perform same transformation. >>> is_same_transform(numpy.identity(4), numpy.identity(4)) True >>> is_same_tra

L2 loss between visible keypoints. \Sum_i [0.5 * vis[i] * (kp_gt[i] - kp_pred[i])^2] / (|vis|)

DELETE ME. Label 0=fake, 1=real. score_real is B x 1, score for real samples score_fake is B x 1, score for fake samples Returns

Assumes that the input mesh {V,F} is perfectly symmetric Splits the mesh along the X-axis, and reorders the mesh s.t. (so this is reflect